Title:
ADVERSARIAL CROSS-MEDIA RETRIEVING METHOD BASED ON RESTRICTED TEXT SPACE
Document Type and Number:
WIPO Patent Application WO/2019/148898
Kind Code:
A1
Abstract:
An adversarial cross-media retrieving method based on restricted text space, comprises: designing a feature extraction network, a feature mapping network and a modal classifier, obtaining a restricted text space by means of learning, extracting image and text features applicable to cross-media retrieval, and implementing the mapping of the image features from an image space to a text space; the diversity of feature distribution between data of different modals is continuously reduced during learning by means of an adversarial training mechanism, thereby achieving cross-media retrieval. The described method is capable of better fitting the behavior performance of humans during a cross-media retrieval task, obtaining image and text features that are more applicable to the cross-media retrieval task, and compensating for the deficiency in expressiveness of pre-training features; and the introduction of an adversarial learning mechanism further improves the accuracy of retrieval by means of the maximum and minimum games between a modal classifier and a feature mapping network.
Inventors:
WANG WENMIN (CN)
YU ZHENG (CN)
WANG RONGGANG (CN)
LI GE (CN)
WANG ZHENYU (CN)
ZHAO HUI (CN)
GAO WEN (CN)
YU ZHENG (CN)
WANG RONGGANG (CN)
LI GE (CN)
WANG ZHENYU (CN)
ZHAO HUI (CN)
GAO WEN (CN)
Application Number:
PCT/CN2018/111327
Publication Date:
August 08, 2019
Filing Date:
October 23, 2018
Export Citation:
Assignee:
UNIV PEKING SHENZHEN GRADUATE SCHOOL (CN)
International Classes:
G06F17/30
Foreign References:
CN108319686A | 2018-07-24 | |||
CN106202413A | 2016-12-07 | |||
CN106095893A | 2016-11-09 | |||
CN104346440A | 2015-02-11 |
Attorney, Agent or Firm:
BEIJING WANXIANGXINYUE INTELLECTUAL PROPERTY OFFICE (CN)
Download PDF: